AI works well for some jobs at a company and badly for others

Companies that treat AI as one tool to roll out everywhere are making a mistake, according to a piece from the Salesforce newsroom. The article is about AI agents, which are programs that carry out tasks on their own. It says they excel at some kinds of work and are unreliable at others, sometimes dangerously so.

Two questions to ask

The article says the uneven results are predictable. Two questions explain where an agent will do well or badly in a business. Has the AI been trained on enough relevant data? And can you verify its answer to make sure it is correct? The answers are not simply yes or no. They fall along a spectrum, and where a task lands helps decide how to split the work between people and AI.

Where AI fits well

Coding is the clear example. It has firm rules and tests that confirm whether the output is right, and that feedback is how models got good at it. The same logic applies to many science and engineering problems, such as finding a material that meets set requirements. Here, the article says, a person's job shifts to deciding what to hand over and setting up checks. Delegate too slowly and rivals may pull ahead. Delegate too early and you can automate mistakes.

Where people hold the missing piece

Some work is easy to check but depends on knowledge the AI never saw. Take getting an order to a customer by Friday. An experienced employee may know that one supplier often runs late, or that another will rush an order if you call the right person. That knowledge often lives in Slack threads and hallway talk, not in the formal process.

The human task is to capture that context and give it to the AI. The article notes that head-mounted cameras now record how service work really gets done, and that Mercor sells expert knowledge to AI labs. Firms that do this well can improve their agents even if rivals use the same AI model. The article adds that companies should keep their own data from training outside models.

Where answers sound right but may not be

Marketing and strategy are different. The AI has plenty of past examples, but nothing can quickly prove its answer right. A campaign may look polished yet feel new without being original. Two rival firms could each get a confident, distinctive-seeming answer and find six months later that their messaging is converging. The article says people should curate, using taste to judge whether something feels right for the company.

A price cut on a struggling product is another case. An agent will answer just as confidently even without the data it needs, and nothing confirms it before you act. The article says that in some cases employees must make and own the decision themselves. The text provided is cut off, so the full fourth category is not covered here.

What to do

• Ask whether the AI has likely seen enough similar work.
• Ask whether you can check the answer quickly and objectively.
• Where you can check, delegate and audit the results.
• Where you cannot, keep a person in charge of the decision.